LLMS ARE GETTING BETTER AT CHARACTER-LEVEL TEXT MANIPULATION

Summary from the TLDR Dev newsletter

LLMs are getting better at character-level text manipulation, such as character replacement, counting, and decoding ciphers. Newer LLMs like GPT-5 and Claude Sonnet 4.5 show improved performance in these tasks compared to older generations, suggesting a better understanding of algorithms beyond simple memorization. This article tests models on Base64 and ROT20 encoded text, showing that larger models are more successful at tasks like decoding and character substitution, even with gibberish text.